Hybrid convolutional neural network–transformer models for liver tumor segmentation: a comprehensive review

Abstract

Liver cancer is a major cause of cancer deaths worldwide, and early and accurate segmentation of liver tumors is a critical step in cancer diagnosis and treatment. However, existing image segmentation techniques have difficulty handling the variability of liver tumors on different image modalities. The emergence of deep learning (DL) and the development of convolutional neural networks (CNNs) have revolutionized image segmentation techniques. However, CNNs have limitations in handling long-range dependencies, which is a critical requirement for tumor segmentation. To overcome these limitations, researchers have proposed hybrid deep learning architectures, which combine CNNs and attention mechanisms or transformers, to integrate local and global information for image segmentation. In this paper, we provide a comprehensive and analytical review of over 50 state-of-the-art deep learning architectures for liver and tumor segmentation. In addition, we provide an extensive evaluation of 38 hybrid and advanced architectures for liver tumor segmentation and a comprehensive discussion of hybrid CNN-transformer architectures. We propose a novel multi-dimensional taxonomy and evaluate the state-of-the-art architectures on various dimensions, including architectural innovation, segmentation accuracy, computational efficiency, and clinical applicability using benchmark datasets such as LiTS and 3DIRCADb. In our critical evaluation of the state-of-the-art architectures, we identify some of the limitations and challenges of existing research and propose a unified evaluation framework and future research directions on self-supervised learning, explainable artificial intelligence (XAI), federated learning, and lightweight architectures.